Download code/THIRD_PARTY_NOTICES.md from zhizeng115/EchoScout-FourTask: direct link, hf CLI and curl.
- Browser
- Download file 1.3 kB
-
https://huggingface.co/datasets/zhizeng115/EchoScout-FourTask/resolve/main/code/THIRD_PARTY_NOTICES.md
- Command line
-
hf download hf://datasets/zhizeng115/EchoScout-FourTask/code/THIRD_PARTY_NOTICES.md
-
curl -L -o THIRD_PARTY_NOTICES.md https://huggingface.co/datasets/zhizeng115/EchoScout-FourTask/resolve/main/code/THIRD_PARTY_NOTICES.md
Third-party components
The source modules omini/models/vision_transformer.py and
omini/models/spatial_ast.py originate from the SpatialAST components in
SLAM-LLM. The latter includes adapted buffer-based spectral operators.
Upstream source: https://github.com/X-LANCE/SLAM-LLM/tree/a68e78f93a12a678438c8b43974fd987abb66927/src/slam_llm/models/SpatialAST
The upstream MIT license is reproduced in licenses/SLAM-LLM-MIT.txt.
Existing public third-party attribution is retained; it is not a statement of
submission authorship.
The binaural frontend follows the BAT / Spatial-AST feature construction. The spectral shims follow torchlibrosa's Fourier and mel-filter mathematics; librosa supplies the mel filters. timm supplies transformer utility layers. Hugging Face Transformers provides the native visual backbones and processors. These installed dependencies and any externally obtained model weights retain their respective licenses. This archive does not redistribute model weights.
The omini namespace is retained solely for source import compatibility. Only
the encoder dependency closure and canonical waveform preprocessing needed by
this package are included; unrelated model families and training services are
excluded. No new blanket license is asserted over third-party code.